An open Hugging Face model with 560B total parameters, 25B active per token and a 1M-token context window. It follows Ling 3.0, which was 127B.
Chinese regulators cleared Alibaba's Qwen AI models for use in Apple Intelligence on iPhones, arriving just ahead of the iPhone Duo launch.
Square launched the Square Kiosk, a large pill-shaped touchscreen ordering device with an agentic conversational voice layer that understands a seller's menu and listens to orders.
A chart from whatmodel.sh, an open-source AI leaderboard, shows Chinese open-weight text models trail US closed frontier models by about seven months, based on 22,113 results across 987 benchmarks from 1,011 sources.
Shopify will let shoppers configure AI agents to purchase products on their behalf, giving more agents access to over a billion listings.
Google Play Store now supports creating apps without developer fees, AI app comparisons and personalized game discovery, letting users vibe-code private apps.
AQuA is a reference implementation that runs beside your production agent on Google Cloud, sweeping logs and traces to catch silent quality regressions. The same roundup covers EmbeddingGemma 2 developer guides, edge multimodal search and TPU video-diffusion kernels.
Xiryn is an open-source personal agent that runs on your own computer: each agent gets its own Linux desktop in Docker, works with any model using your keys, has an encrypted vault, and is controllable via Telegram.
Nvidia's RTX Spark superchip combines gaming PC, creative studio and personal AI supercomputer roles, with up to 128GB unified memory and 1 petaflop of AI compute. DLSS 5 updates ship alongside.
Google released Gemini 4 Argon, its next flagship model, with safety work, robotics applications, and a demo of optimizing a 200,000-line Python codebase overnight.
Alibaba opened beta sign-ups for Qwen Go, a personal agent that files flight-delay claims and Taobao refunds, and launched N1 Pro and N1 AI glasses starting at RMB 2,218 (about $450).
Apigee hybrid 1.17.0 adds native Model Context Protocol support, letting the MCP data plane run inside a user's own Kubernetes cluster so AI agent traffic and tool execution stay on their infrastructure.
Parsimony is an open-source benchmark that measures how much extra code AI models add beyond what is needed. It uses patches from existing benchmarks like DeepSWE to penalize excess code in long-term projects.
Meta releases Muse Wonder and Muse Wonder Pro, two models from its revamped Superintelligence Labs that it claims rival the industry's best, following Muse Spark and its Muse agent.
Physical Intelligence launched an API beta it calls frontier models for every robot, letting developers access its general robot foundation models remotely. Pricing and availability details were not in the source text.
Security incident at a leading AI lab raises concerns over insider threats and IP protection.
An OpenAI employee was arrested and charged with using a hidden recording device in a San Francisco office, per the company and the US Justice Department.
A large vendor-funded push into agent safety research, shaping UK AI policy and evaluation work.
Nvidia will provide UK researchers about $50M in computing annually, totaling $250M over five years, for AI agentic safety and security research.
Shows current AI models remain weak at automating chip design, setting a baseline for progress.
Nvidia and ChipAgents released a benchmark measuring how AI models perform at chip design tasks. Anthropic's Fable 5.1 powered the best agent, scoring only 16.6%.
Google's new flagship model pushes frontier reasoning, coding and robotics capabilities.
Google released Gemini 4 Argon, its next flagship model, with safety work, robotics applications, and a demo of optimizing a 200,000-line Python codebase overnight.
EUV lithography supply is a hard bottleneck on global AI compute expansion.
ASML raised its full-year sales outlook and forecast Q4 revenue above expectations as AI-driven chip demand surges, but warned it will be unable to meet customer demand in 2027 and 2028.
Leadership change over OpenAI's data center buildout signals shifting infrastructure strategy.
OpenAI named former xAI executive Uday Ruddarraju to lead its infrastructure and compute units. Current computing chief Sachin Katti will depart next month, per WSJ.
Extends agentic payment infrastructure onto a major public blockchain via embedded wallets.
Ripple is working with Stripe-owned Privy to embed wallets on the XRP Ledger and enable AI agent auto-payments, with AI payments on XRPL reportedly passing 13-14 million.
Signals escalating legal conflict over talent and trade secrets between OpenAI and Apple.
Business Insider reports OpenAI stated it is very concerned its trade secrets could help Apple's push into AI-enabled devices, intensifying the two companies' legal dispute.
First high-profile criminal probe of robot conduct raises accountability questions for deployed humanoids.
Police in Cologne, Germany seized a humanoid robot after it allegedly sexually harassed a 16-year-old girl on the street, prompting an investigation into its owners.
Flags a novel security risk as capable models proliferate onto everyday consumer hardware.
Anthropic warns small AI models could act as worms, self-replicating across smart home appliances like lawnmowers and TVs within months, without needing data centers.
Links a major life-sciences tools maker with Nvidia to automate scientific experimentation.
Danaher is collaborating with Nvidia to integrate AI into scientific instruments, with initial work focused on select instruments to advance autonomous science and AI-supported research.
The CEO of the top AI chip supplier setting safety conditions on sales could shape industry norms.
Jensen Huang said he would halt chip supply to any customer developing unsafe AI software, while adding he thinks current concern about the technology is overblown.
A new frontier speech-to-speech model powers real-time voice agents for consumer and enterprise use.
xAI released Grok Voice 3.0, a speech-to-speech model for building voice agents; it is reportedly already handling Starlink and Tesla calls.
Agent payment volumes on x402 signal growing but unverified adoption of machine-to-machine payments.
Coinbase launched an agent payment stack on the first anniversary of its business unit. Reports cite 138-200 million x402 payments by AI agents, though some may be fake.
Brings petaflop local AI compute to consumer laptops, enabling on-device models and creation.
Nvidia's RTX Spark superchip combines gaming PC, creative studio and personal AI supercomputer roles, with up to 128GB unified memory and 1 petaflop of AI compute. DLSS 5 updates ship alongside.
A European humanoid maker building its own robot foundation model signals vertical integration in physical AI.
German robotics unicorn NEURA Robotics introduced Neuron AI, a multimodal foundation model that gives robots perception and memory for physical AI and humanoid applications.
Signals heavy infrastructure investment push across Europe.
Nvidia's Huang plans roughly $2 billion in European AI deals in the coming week, as French companies deploy agentic computing across space, financial services, the public sector and research.
Clears the path for Apple Intelligence in China and cements Alibaba as the local AI partner for a major device launch.
Chinese regulators cleared Alibaba's Qwen AI models for use in Apple Intelligence on iPhones, arriving just ahead of the iPhone Duo launch.
Major retailer adopting agentic commerce shows mainstream uptake of AI-driven shopping.
Target is expanding agentic commerce efforts to open new digital growth avenues and upgrading its employee app for more efficient stores, analysts and Retail Dive report.
Sovereign AI for core banking reflects banks' push to keep regulated workloads on local infrastructure.
French software firm SBS launched what it calls the world's first sovereign AI platform built for the banking market, aimed at regulated financial institutions.
HiRO-ACEv2 is an AI weather model that simulates extreme weather at 3 km resolution for a fraction of the usual cost, released on Hugging Face.
An open Hugging Face model with 560B total parameters, 25B active per token and a 1M-token context window. It follows Ling 3.0, which was 127B.
Agent-Reflex is a training framework that teaches language agents to reflect on noisy MCP tool feedback and user corrections, improving failure recovery in long-horizon tasks.
GreenWAM is a world foundation model that both predicts how a robot's actions change a scene and chooses those actions, aiming to work across different robot bodies.
Paper Tev presents a method to scale long-horizon agentic training environments using dense progress rewards, aiming to help open models work autonomously for hours to days.
A VAE that reduces checkerboard artifacts in Qwen Image 2.1 generations, shared on Hugging Face. Differences are only visible when zooming past 500%.
A pixel-art LoRA for image models Krea2 and Anima, downloadable from Civitai with the trigger word 'aziib pixel style'.
Urbex is a research framework that lets a vision-language agent actively search a city-scale 3D scene, querying landmarks and zooming instead of ranking precomputed instance proposals.
VideoJev is a family of models that turn video understanding and interaction into structured decision tasks, adapting pretrained MLLMs with a shared prediction head instead of autoregressive text answers.
Jeva is a paper arguing current system-one model designs miss four structural requirements, including distinguishing answer absence from uncertainty and permutation invariance, and proposes models built around them.
A research paper proposing lightweight latent future prediction to connect vision-language-action policies with world action models, aiming to model action consequences without costly video generation.
A paper examines how GUI agents should choose the modality for retrieved memory, arguing existing agents use a fixed representation instead of adapting the format to the task.
A research paper on sparse-to-dense motion completion: recovering dense mesh motion from sparse vertex displacements. It shows residual priors trained on one geometry family often fail to beat parameter-free harmonic extension on other families.
An AI agent project decomp.wtf runs locally to decompile games you own into C that compiles to identical bytes. Developer Popcar describes a wave of AI decompilation enabling mashups and preservation projects.
ComfyUI, the node-based GUI and backend for local diffusion model inference, hits v0.40.0 with CSP changes allowing blob: and data:, plus comfy_attention support across more models.
Xiryn is an open-source personal agent that runs on your own computer: each agent gets its own Linux desktop in Docker, works with any model using your keys, has an encrypted vault, and is controllable via Telegram.
mailcheap is an affordable open-source newsletter tool for the agentic era; you can self-host it or use the hosted mailcheap.ai service. It hit 918 stars in its first week.
pydantic-ai is a Python framework for building AI agents, voice, and image apps across models; v2.56.0 adds layered project and personal skill libraries and makes browser-use telemetry opt-in.
Scout is an open-source browser agent that automates web tasks in a browser, mirrored from Outcrawl's own repository; it gained 165 stars in its first week.
A Blender add-on lets you create and edit 3D models by chatting with AI: send text or an image and get working bpy code. Open-source, ~211 stars in week one.
diff-tour is a read-only terminal UI that walks you through a git diff the way a senior engineer reads it, turning a change into a guided tour. Written in Go.
resimind is an open-source Python project providing neuro-symbolic agents with verified memory: MCP math tools, an algebra repair skill, and an open challenge so agents can verify solutions and reuse checked rules.
new-term-radar is a Python tool that scans X, Hugging Face, GitHub and sitemaps daily to catch names of new AI models, products and repos before keyword tools list them. ~891 stars in week one.
Parsimony is an open-source benchmark that measures how much extra code AI models add beyond what is needed. It uses patches from existing benchmarks like DeepSWE to penalize excess code in long-term projects.
Hindsight 1.0 and Memlake are open-source memory tools for AI agents, letting agents store and learn from past experience. Launched on Product Hunt.
Gausify Orbit is a browser tool that turns AI-generated 360 orbit videos into Gaussian Splats, now released under the MIT license with new updates.
Tencent released Agent Memory, an open-source memory layer for AI agents that runs entirely locally on SQLite, letting developers add persistent recall to agents without a cloud service.
Anima is an open-source Claude Code skill that builds interactive scientific figures instead of text explanations, letting users visualize biological processes on request.
A JavaScript tool that ferries OpenCode's free models to WorkBuddy, offering an isolated runtime, concurrency checks, decision caching, one-click import and a local OpenAI-compatible forwarding port.
An open-source Python MCP server and agent skill that pulls McDonald's China menus and coupons and computes the cheapest meal combination, e.g. a Big Mac combo at ¥21.9 versus ¥47.
Alibaba opened beta sign-ups for Qwen Go, a personal agent that files flight-delay claims and Taobao refunds, and launched N1 Pro and N1 AI glasses starting at RMB 2,218 (about $450).
Android is rolling out scam defense features, including an enhanced protection setting that uses Gemini to warn users about scammy messages in Google Messages and on calls. The feature rolls out as opt-in protection.
YouTube is adding a 'Made with AI' label to tracks that were altered or fully generated with AI, mirroring its existing AI labeling for videos.
Google Play Store now supports creating apps without developer fees, AI app comparisons and personalized game discovery, letting users vibe-code private apps.
xAI launches an X subscription combining Grok, the Grok Bot agent, Cursor and X Premium, with docs covering Grok Bot security and FAQ.
Deezer said a toggle letting listeners hide fully AI-made or AI-generated music from search and playlists will roll out by the end of the year.
L-Acoustics released new software that uses AI to isolate a singer's voice so it cuts through concert noise. Céline Dion and Harry Styles reportedly use the technology.
A report compares the digital closet from the film 'Clueless' to new ChatGPT shopping and styling features that let users assemble outfits with AI assistance.
Users can add and manage topics for their Facebook and Reels feeds; Meta says recommendations respond right away to interested or not interested taps.
Google Workspace will automatically connect Google Tasks to Gemini AI starting in November, letting Gemini surface and act on to-dos.
Rocket Money's Rowan is a financial AI assistant users can text for money help, according to a 9to5Google hands-on.
Higgsfield now makes AI videos longer than 10 minutes with GPT-6 Astra, sharper visuals via GPT-Image-2.5, and voice filmmaking with GPT-Live.
A how-to explaining how to set up and use Gemini Auto Browse in Chrome, including available tasks and its current limitations.
Perfect Corp is partnering with Laboratoires Lierac to bring AI-powered skin diagnosis to consumers in Europe, using the beauty brand's products and Perfect Corp's AI skin analysis technology.
OpenAI removed UTM tracking parameters from links surfaced in ChatGPT results, a change affecting how publishers and marketers see referral traffic from the chatbot.
South Korea's nationwide 'Everyone's AI' service faces its first public test on Oct. 16 with SK Telecom, KT and Kakao participating.
Microsoft is responding to a wave of AI decompilation and recompilation projects that let players port classic games to new devices or merge code from two titles.
Mashable reports Google is changing Gemini access for free users, with implications for what features remain available without a paid plan.
Android Authority reports Google made it easier to assign tasks to Gemini, streamlining how users hand off actions to the assistant on Android devices.
Meta's assistant app adds daily briefings, reminders, email and calendar connections, and in-depth research reports.
YouTube's contextual AI search, branded Ask YouTube, is rolling out to all users on Android and iOS in the US, no longer limited to YouTube Premium. It answers conversational queries about videos.
Google's Gemini app redesign integrates Spark into its Task and scheduled page, with a rollout speed bump for some Gemini Skills users.
Pulse Local is a Mac budgeting app (Apple Silicon, beta) that answers questions about your spending using models running on your machine. Local mode needs no account; the ledger and questions stay on-device with no analytics or crash logs.
Microsoft's October Windows 11 update adds File Explorer tweaks and Copilot key remapping in a new push for its Copilot assistant.
A Dutch person gave Claude six photos of a pothole; it rebuilt the hole in 3D and simulated a 30 km/h hit, prompting city hall to fix it in two days. Rockefeller Foundation also teams with OpenAI on potholes.
New releases: Visual Studio Code 1.141.1, an open-source code editor for macOS, Windows and Linux; plus Cursor 3.24.18, Zed 1.23.3, Warp 0.2026.09.30.08.29.stable_01, Kiro 1.3.0 and Trae 2.3.61406.
Anthropic describes redesigning its source control system after its monorepo hit 30,000+ merges in a day, arguing coding is no longer the bottleneck now that Claude produces large volumes of code.
Apigee hybrid 1.17.0 adds native Model Context Protocol support, letting the MCP data plane run inside a user's own Kubernetes cluster so AI agent traffic and tool execution stay on their infrastructure.
AQuA is a reference implementation that runs beside your production agent on Google Cloud, sweeping logs and traces to catch silent quality regressions. The same roundup covers EmbeddingGemma 2 developer guides, edge multimodal search and TPU video-diffusion kernels.
Nvidia open-sourced Run:ai as part of its DSX AI factory platform; Nebius made it available on its AI cloud and vCluster Labs turned it into a managed cluster product. Santander's Botín also said AI will change banking jobs.
OpenAI details how Amp builds AI agents that work beyond coding, using GPT-6 Astra, Ultrafast, GPT-Live-1 and Sign in with ChatGPT.
Antigravity, an agent orchestration platform for macOS, Windows and Linux, released version 2.23.0, which adds an SDK with local model support for hybrid agent orchestration.
Floqer is an autonomous customer knowledge platform giving B2B sales and marketing teams an AI-powered database that continuously collects, organizes and updates customer information, replacing manual prospect research.
Epic Games previews a faster, Linux-supporting PC games launcher it says was 'vibe coded' with AI-assisted development, part of its public AI coding push.
Google now lets anyone distribute vibe-coded Android apps with just a Google account and run private personal apps, plus a Linux terminal for AI app building.
Open-source Gemini CLI 0.64.0, an AI terminal agent, adds A2A server V1-to-V2 settings migration, ACP PromptResponse usage bridging and usage_update notifications.
Quadient launched an AI agent that automates customer account management and email handling for businesses.
Zed is a multiplayer code editor for macOS, Windows and Linux. Version 1.24.2 is a minor maintenance release.
SD Times reports on a release gate approach for catching and blocking AI agent incidents before deployment, adding checks to agent release pipelines.
AWS published a guide on building AI agents that browse and complete transactions using Anchor Browser with Bedrock AgentCore payments.
Rival.io says it has eliminated token charges for developers building with open-weight AI models, removing per-token billing from its platform. Details on pricing tiers and supported models are limited in the source.
An AWS guide shows building agents on Bedrock AgentCore that automate PowerPoint creation from enterprise documents, using a supervisor agent, RAG and Guardrails to generate grounded decks in under 90 seconds.
Tamnoon built Tami, an agentic remediation orchestrator on Amazon Bedrock that investigates and safely fixes cloud security findings from CNAPP and cloud-native sources, cutting critical exposure over 90% in 90 days.
Precisely launched an MCP Server for Precisely Automate, connecting AI assistants to SAP automation.
NeuBird's Agentic Reliability Center (ARC) is a platform for monitoring and governing AI agents running in production environments, aiming to keep agent behavior reliable and accountable.
Omega Point's MCP server is now Anthropic verified in Claude's marketplace, letting Claude users connect to Omega Point's compute capabilities through the Model Context Protocol.
Conductr is an agent runtime that unifies deployment, data and observability for AI agents in one platform, aimed at teams running agents in production. Railtown AI announced its launch.
Pi 1.0 is an open-source coding agent that its developers say cuts token use by 40% versus comparable agents.
Anthropic released materials on using Claude Code for data engineering plus a CISO guide to deploying Claude and a rebroadcast on monitoring and securing agents on Google Cloud.
Scuba, founded by Slack veterans, launched with $34 million to build a shared context engine giving teams and their AI agents common memory and state. It targets enterprise teamwork and agent workflows.
